Researchers have unveiled TimeBraid, a family of unified models that combine pretrained language models with pretrained time-series foundation models. The approach aligns a Qwen3 LLM and a TimesFM-based time-series model through interleaved global residual attention layers, allowing both modalities to share a single representation space where they can be understood and generated together.
The core idea is that language models possess broad world knowledge but cannot perceive continuous signals, while time-series foundation models understand signals but cannot process text. TimeBraid preserves both capabilities: series remain as continuous patches, text remains as tokens, and both are woven onto one timeline using shared causal attention.
The training recipe combines a unified prompting scheme for diverse time-series and text tasks, stabilized joint training, and supervision from 2.2 million curated series–text pairs and 4.9 million instruction-tuning samples. According to the authors, the resulting models inherit instruction following and reasoning from the language side, and continuous-signal perception and zero-shot forecasting from the time-series side.
Across 12 benchmarks covering time-series perception, understanding, reasoning, and both context-aided and unimodal forecasting, TimeBraid remains competitive with far larger general-purpose models and task-specific counterparts. Notable results include the top average rank among 20 models on Time-MMD contextual forecasting across 9 domains, the best macro MSE/MAE and lowest CRPS on CGTSF and CAF (with GPT-5.4 included on CAF), and 36.6% on TemporalBench for TimeBraid-6.7B versus 38.5% for GPT-5.4 — described as on par and ahead of Claude Sonnet 4, Gemini 2.5 Flash, and GPT-4o. On TimeSeriesExam, scores rose from 50.1 to 60.1 to 63.1 as model size scaled from 1.2B to 2.5B to 6.7B. On Ctrl-F, TimeBraid is described as the only non-LLM model whose forecasts follow the text condition.
The TimeBraid-2.5B weights and inference code are released under the Apache-2.0 license, with a project page, GitHub repository, Hugging Face model card, and arXiv paper all publicly available.